Dynamic path planning in autonomous driving

With the development of emerging technologies such as the Internet of Things (IoT), Artificial Intelligence (AI) and fifth-generation cellular network technology (5G), good conditions are provided for the level of automation, providing the necessary assistance to coordinate the environment, planning decisions and path optimisation for self-driving vehicles. This paper briefly outlines and analyses the key technologies required to realize autonomous driving and summarizes the corresponding research results for each part; by considering the operation process of self-driving vehicles, this paper chooses path planning as the focus of research and presents it in the form of a methodology; finally, it concludes with a summary of personal opinions and a discussion of possible future research directions.


Introduction
The concept of "automation" is relative to human activity, through the rational design and use of control devices, so that the controlled unit (this article refers to machinery and institutions) in part of the scene in an orderly and efficient way to achieve the purpose, in order to get rid of human supervision of the whole process, and to liberate the workforce.Automation technology is a comprehensive technology with close ties to the fields of control theory, information theory, systems engineering, computer technology, electronics, hydraulic and pneumatic technology, automatic control, etc. "Control theory" and "computer technology" have the greatest influence on the development of automation technology.In this paper, the superiority of "automation" is illustrated by "scene understanding", "motion planning" and "vehicle control", which are the essential components for autonomous driving.
"Autonomous driving" has been hailed as the disruptive innovation of the next few years.As a technology-driven programme, limited by the level of technology and the needs of society, the major milestones in the field of autonomous driving have been taking shape since the middle of the last century, as shown in Figure 1 below.According to a recent report by the World Health Organization (WHO), road injuries are among the leading causes of death worldwide.Approximately 1,300,000 people are killed annually in motor vehicle accidents.More than half of this alarming number consists of vulnerable road users (VRU), such as pedestrians and non-motorized users [2].As a result, how to regulate the behaviour of both public and private vehicles on the road has become a hotly debated topic.
The intelligent transport systems (ITS), which rely on self-driving vehicles, are gaining more and more attention in order to reduce human-generated traffic accidents and further ensure traffic safety; and, with the rise of the Internet of Things (IoT), Artificial Intelligence (AI) and fifth-generation cellular network technology (5G), provide the necessary assistance for autonomous vehicles to coordinate their environment, make decisions and optimise paths.It is expected that self-driving cars will account for 25 percent of the global private car market by 2040 [3].
Although "self-driving cars" are not yet widely available on the market, a number of companies around the world are already testing and developing them in real-life environments.For example Startup AutoX and US company Waymo are offering self-driving taxi testing services to a small public in Shanghai, China and New York and San Francisco respectively; Rio Tinto Mining in Western Australia is currently operating the world's largest fleet of self-driving transport vehicles, etc.
The real-world road environment is complex and dynamic, and the driving environment can be impacted by traffic regulations, traffic participants, and the vehicle's physical limitations, making it difficult to achieve safe and effective autonomous driving [4].This paper will investigate key technologies for autonomous vehicles, analyse the vehicle's state of motion, and conclude how to optimise the vehicle's overall performance.

Related work
The Society of Automotive Engineers (SAE) has proposed a standard J3016 that classifies self-driving cars into six classes, as shown in figure 2   In general, Level 0-2 vehicles rely heavily on the human driver to make decisions; whereas Level 3 vehicles make certain decisions autonomously, such as maintaining a safe distance and following the vehicle in front of them; Level 4 vehicles are capable of fully autonomous driving modes under specific conditions, but the driver can still operate manually in the event of a system failure; and Level 5 vehicles are fully automated drivers [5].
The subject of this paper is biased toward autonomous vehicles that can independently perform complete decision-making behavior in most situations.The process is usually divided into three stepsscenario understanding, motion planning, and vehicle control -and special algorithms are used for each.

Scene understanding
The essence of scene understanding lies in the transmission of driving-accurate road information (e.g.lanes, signals, pedestrians, traffic signs, etc.) via different sensors (e.g.LIDAR, cameras) around the vehicle.
Balado et al. propose to first acquire a point cloud sample of the road environment by mobile laser scanning (MLS) and then use PointNet to semantically segment the sample to identify the points that make up each object (road surface, guardrail, boundary) [6].
Zhang et al. introduced a framework for detecting the spatial position of 3D targets (MCK-NET), capable of mining depth information from monocular images, providing ideas for reducing the use of 3D sensors and providing a relatively low-cost solution with monocular cameras [7].

Movement planning
In recent years, several research on dynamic route planning have been done.These research may be split into four basic categories: grid-based methods, potential field methods, sample-based methods, and discrete optimisation methods [8].

Grid-based method.
The environment is mapped into a series of cells, some of which represent obstacles at that place in the environment, and optimum search algorithms such as A* and D* are commonly used to determine the best path from the original site to the destination location while avoiding obstacles.Nonetheless, the gradual nature of search algorithms might result in an exponential rise in computer complexity.Grid-based methods are often appropriate for low-speed path planning [8].

Potential field methods.
First is the potential field method.The central idea of the artificial potential field method is to construct a smoothing function in the bit-shaped space that has a higher value near the obstacle region and a lower value in the opposite direction, with the lowest value at the target state, and then to use a simple gradient descent method to easily plan a path from any initial state to the target state [9].

Sample based methods.
Second is the sample-based method.This method constructs a collisionfree path from the initial position to the destination by sampling configurations that describe the vehicle's position and orientation.This method is suitable for planning in high-dimensional spaces.No explicit construction of obstacle regions or exploration of a large number of bit-shaped spaces is required, resulting in a significant reduction in computational effort.Rapid Exploration Random Tree (RRT) algorithm and its variants are commonly employed for nonholonomic path planning [8].

Discrete optimisation methods.
Third is the discrete optimisation method.This method generates a mathematical model containing a cost function and constraints to facilitate continuous optimisation outcomes.In contrast to the sampling method, discrete optimisation results in a final trajectory that is highly flexible to environmental changes.In contrast to sampling methods, it facilitates proactive action to overcome complex situations.Unfortunately, this process requires significant computational resources to solve spatio-temporal planning problems with dynamic obstacles, the solver may not produce results quickly, and in practical situations, alternate trajectories should be generated in a timely manner [10].

Vehicle control
Vehicle control is a closed-loop process involving the transfer of sensor data to the vehicle control unit and the adaptation of algorithms to operate the vehicle actuation terminals.
Yu et al. optimised a multi-person game theoretic model to fully consider multiple scenarios for whether to perform a lane change in an environment containing both autonomous and human-driven vehicles (AV-HV) [11].
Yang et al. presented a hybrid motion planning framework (HMPF) to enhance the performance of motion planning, which comprises of learning-based behavioural planning and optimisation-based trajectory planning.The former makes use of Deep Reinforcement Learning (DRL) algorithms to create behavioural decision instructions based on environment-aware information, learning from interactions between self-vehicles (EVs) and other human-driven vehicles (HDVs).Instead of the Cartesian coordinate system used in conventional motion analysis, Frenet coordinates based on roads are employed in the latter, substantially decreasing the computer power necessary while providing a logical route [12].The framework is presented in figure 3    Liu et al. proposed to obtain motion data from real scenes at the spatial level and construct an exclusive relative area velocity vector (EARVV) algorithm to address the effect of coupled vehicle motion on collision avoidance and to provide a spatial region for collision avoidance trajectories; at the temporal level, the changes of the object vehicle and its relative velocity vector are fully considered to define the spatio-temporal collision avoidance trajectory [14].The theoretical underpinnings of the recommended approach is shown below in figure 5.

Figure 5.
Theoretical underpinnings of the recommended approach for avoiding catastrophic collisions [14].

Conclusions
Based on the three steps (scene understanding, motion planning, and vehicle control) described above for implementing dynamic path planning for self-driving vehicles: Zhang et al. mentioned making full use of monocular images to reduce the equipment cost of 3D cameras, but pixel distortion cannot be ruled out in bad weather when devices such as LIDAR remain indispensable.Also, the sensor devices are limited in the amount of road information they can store, and I think that for the time being self-driving vehicles will be developed and put into use first on parts of public roads that are consistently and frequently used over time.In the future, if self-driving cars become the overwhelming choice for travel, it may be easier for vehicles to locate and exchange information with each other.
Currently, a combination of artificial potential field methods and discrete optimisation methods are often used to test virtual driving scenarios built from existing data.The limitation of the analytical approach lies in the balance between calculating a sufficient decision method and reducing computation time.
As for vehicle control, in conditions where fully autonomous driving is not yet mature, we have to ensure a certain degree of operability of the vehicle for the passengers during the test phase; and given that the vehicle's decisions are not necessarily optimal each time, the vehicle actuators need to be matched to the corresponding algorithms and the vehicle itself needs to ensure its range under autonomous driving.
There is still a long way to go before full driving automation is achieved, and some of the experimental data from current research in this area are still based on the recording of real vehicle data and scenario simulation by computer software.The main challenge for self-driving cars is to have a better understanding of the complex real world.In addition to the methods described above, deep reinforcement learning and neural structured networks have shown significant performance in characterizing objects in the traffic environment and accumulating practical experience respectively, and there is no best method but only continuous optimisation and progression.

Figure 1 .
Figure 1.Self-driving technology timeline [1].According to a recent report by the World Health Organization (WHO), road injuries are among the leading causes of death worldwide.Approximately 1,300,000 people are killed annually in motor vehicle accidents.More than half of this alarming number consists of vulnerable road users (VRU), such as pedestrians and non-motorized users[2].As a result, how to regulate the behaviour of both public and private vehicles on the road has become a hotly debated topic.The intelligent transport systems (ITS), which rely on self-driving vehicles, are gaining more and more attention in order to reduce human-generated traffic accidents and further ensure traffic safety; and, with the rise of the Internet of Things (IoT), Artificial Intelligence (AI) and fifth-generation cellular network technology (5G), provide the necessary assistance for autonomous vehicles to coordinate their environment, make decisions and optimise paths.It is expected that self-driving cars will account for 25 percent of the global private car market by 2040[3].Although "self-driving cars" are not yet widely available on the market, a number of companies around the world are already testing and developing them in real-life environments.For example Startup AutoX and US company Waymo are offering self-driving taxi testing services to a small public in Shanghai, China and New York and San Francisco respectively; Rio Tinto Mining in Western Australia is currently operating the world's largest fleet of self-driving transport vehicles, etc.The real-world road environment is complex and dynamic, and the driving environment can be impacted by traffic regulations, traffic participants, and the vehicle's physical limitations, making it difficult to achieve safe and effective autonomous driving[4].This paper will investigate key technologies for autonomous vehicles, analyse the vehicle's state of motion, and conclude how to optimise the vehicle's overall performance. below:

Figure 2 .
Figure2.SAE J3016 standard[5].In general, Level 0-2 vehicles rely heavily on the human driver to make decisions; whereas Level 3 vehicles make certain decisions autonomously, such as maintaining a safe distance and following the below.

Figure 3 .
Figure 3. Hybrid motion planning framework[12].Guo et al. suggested a map-enhanced generative adversarial network (ME-GAN) for vehicle trajectory prediction in order to make the most of the information included in high-definition (HD) maps [13].The netwok is illustrated in figure 4 below.